Reproduce “Non-asymptotic error bounds for probability flow ODEs under weak log-concavity” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gitte Kremling, Francesco Iafrate, Mahsa Taheri, Johannes Lederer
Researchers from the University of Hamburg established non-asymptotic 2-Wasserstein convergence bounds for probability flow ODEs in score-based generative models, by relaxing the target data distribution assumption from strong to weak log-concavity. The work rigorously quantifies errors from initialization, score approximation, and numerical discretization, showing that the asymptotic error rates remain consistent with prior findings despite these weaker assumptions.
Key insights
- Weak log-concavity and Lipschitz continuity of the score function are preserved and propagated through the forward SDE, ultimately leading to strong log-concavity of intermediate distributions after a finite "regime shift" time.
- The total error in 2-Wasserstein distance can be precisely bounded by the sum of explicit, non-asymptotic components for initialization, discretization, and score approximation errors.
- Despite a substantial relaxation of data assumptions from strong to weak log-concavity, the asymptotic rates of the error bounds with respect to diffusion time, step size, and score-matching error remain identical to those derived under stronger conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.